A Survey of Color for Computer Graphics

A Survey of Color for Computer Graphics SIGGRAPH 2001
face reflectances of non-fluorescent objects 2 are often very smooth, as is the spectral distribution
of daylight or of incandescent light. Therefore it is possible to adapt the
sampling to give a more efficient representation. Gary Meyer, Roy Hall and Andrew
Glassner have all proposed methods for doing this, and Roy Hall has a paper surveying
this topic in the July 1999 issue of IEEE Computer Graphics and Applications. The most
efficient representations use as few as 4 samples/spectrum. Note, however, that a more
concise representation may mean a more expensive reconstruction algorithm when the
spectra are actually applied in the rendering.
Another way to create an efficient representation for a set of curves is to define basis
functions that can be linearly combined. With this approach, only the scalar multipliers
for the basis functions need to be stored for each color. Maloney and Wandell have applied
this technique to a variety of surface datasets, and Mark Peercy has applied this
approach to computer rendering in his SIGGRAPH ’93 article.
To display spectra on a monitor, they must be converted to RGB monitor values. Using
trichromatic theory and the CIE color matching functions, it is straightforward to convert
from spectra to tristimulus values. Because a digital monitor is also a three dimensional
additive color system, there is a linear transformation that converts from CIE tristimulus
values to the intensity needed for each of R, G and B. This matrix is derived from the tristimulus
values that describe the monitor primaries, as is shown in the slides.
RGB representations evolved by simply applying grayscale shading models to the red,
green and blue image components independently. The effect is that any color is described
as a linear combination of the three display primaries. If all we want to do is specify individual
colors, or additive mixtures of them, this is fine. However, in rendering, we
multiply the spectrum of the light by the reflectance of the surface. This is not guaranteed
to produce another spectrum that can be represented as a linear combination of the display
primaries. Said another way, if the RGB values are the tristimulus values for the
spectra, the product of the tristimulus values does not equal the tristimulus values of the
product of the spectra. The exception is when one of the spectra is an equal energy white,
which is often the case for simple graphics systems. A SIGGRAPH ’91 paper by Carlos
Borges provides an analysis and bound for the error created by using an RGB model.
Using an RGB representation for color can be very effective. Its failings only really appear
as the lighting and shading models become more complex. The difficulty of
constructing a simple case where RGB fails can be seen using the demonstrations in the
Color Playground from Brown University that are included with the CDROM version of
these notes.
Viewing Rendered Graphics
All graphics rendering algorithms must ultimately produce an image that can be displayed
on a monitor or some other digital display device. The rendering color space is an
unbounded, linear intensity space that must be mapped to the bounded, non-linear color
space of a display system. As in image reproduction, the issues include definition of the
correct RGB primaries, intensity mapping, out-of-gamut colors, and appearance consid-
2
Phosphors, which are used in CRT displays or in the coating of fluorescent lights, contain high-energy
spikes that must be carefully sampled.
Maureen C. Stone 18
StoneSoup Consulting